This week felt like a quiet shift in the underlying assumptions of local AI development. Google decoupled its agent runtime from the IDE, Apple Silicon got meaningfully faster for open model inference, and a new specification took a real swing at the fragmented mess of agent tool discovery. Here's what's worth your attention.
Google ships Antigravity 2.0 agent desktop app
Antigravity's agent surface has been extracted from the IDE and shipped as a standalone desktop app for macOS, Linux, and Windows. The new version adds dynamic subagents, async task management, JSON hooks for external integrations, and cron-based scheduled execution.
The IDE coupling was always the wrong abstraction for non-coding agent work. If you're running agents against docs, data pipelines, or internal tooling, you don't want IDE context bleeding into your agent state. Async task management and native scheduling mean you can now fire off recurring agent jobs without babysitting them or duct-taping cron to a shell script that pings an API.
Slash commands — /goal, /schedule, /grill-me — expose the core capabilities immediately, which is a reasonable first-run experience. The IDE remains available during the transition, so there's no forced cutover.
Verdict: Evaluate. If you're using agents for anything beyond code generation, this is worth an hour of your time this week. The scheduling and async primitives alone change what recurring agent workflows look like.
Nativ runs open models locally on Apple Silicon
Nativ is an MIT-licensed desktop app that runs open LLMs directly on M-series Macs via MLX-VLM. No API keys, no cloud round-trips, no cold-start latency. It exposes real-time telemetry — tokens/sec, memory usage, thermal state — and serves a local endpoint your coding agents can point at.
The telemetry story is underrated here. Most local inference tools give you a model and a chat window; Nativ gives you the instrumentation to actually understand what's happening on your hardware. That matters when you're tuning which model fits your RAM envelope or debugging why your M1 Pro is thermal-throttling mid-generation. The endpoint integration means you can swap out cloud LLM calls in your agent config and keep prompts off external servers entirely.
The constraint is straightforward: model size is bounded by device RAM. An M1 with 16GB has a real ceiling. Image, video, and audio modalities are supported via the MLX-VLM backend, which makes this more capable than most local runners at this stage.
Verdict: Ship — if you're on M-series hardware and have been meaning to cut cloud dependencies for local dev workflows. Zero onboarding friction, open-source, works today.
ARD specification standardizes agent capability discovery
Agent Resource Discovery (ARD) proposes a standard: publish an ai-catalog.json at a well-known path on your domain, and agents can discover and verify your tools via federated registries using cryptographic identity tied to domain ownership.
The problem it's solving is real. Right now, multi-org agent systems either hardcode integrations or maintain bespoke internal registries that don't compose across organizational boundaries. ARD gives you a portable, interoperable layer — if both orgs publish catalogs, agents can locate and trust capabilities at runtime without anyone writing a custom adapter.
The specification is live, there's a quickstart, and it's Apache 2.0 licensed. The production blocker is honest: full Gemini Agent Platform integration is still "coming months." The infrastructure to host the registry side requires some setup, but the entry point is genuinely low-friction.
Verdict: Evaluate. If you operate multi-org agent systems, stand up a catalog and get familiar with the spec now. You don't want to be retrofitting this when platform support lands.
Gemma 4 generates tokens 90% faster on Apple Silicon
Ollama's Gemma 4 implementation now uses multi-token prediction with auto-tuning across MLX and llama.cpp backends. The result is roughly a 90% token generation speedup on M-series chips. No configuration changes, no model retraining, no prompt engineering required — update Ollama and the speedup is just there.
This crosses a threshold. Local LLM inference on Apple Silicon has been viable for experimentation but awkward for coding agents that need low-latency completions. A 90% speedup on generation means the iteration loop is now tight enough for real-time use cases that were previously impractical. The zero-config nature of it matters: existing Ollama workflows get faster automatically on v0.31.1+.
Verdict: Ship. Run ollama --version, update if you're below v0.31.1, done. There's no reason to defer this one.
Google AI Edge Gallery adds MCP, scheduled notifications
Google's AI Edge Gallery now supports Model Context Protocol over HTTP for mobile agents, alongside on-device task scheduling and persistent chat history. Gemma 4 on-device can now call external tools — calendar, email, web — without server round-trips, with local inference handling the reasoning.
The MCP integration is the interesting piece. Wiring a mobile agent to real data sources while keeping inference local changes the privacy and latency profile of edge agent workflows significantly. The practical constraint is context window size: smaller on-device models mean you need to keep MCP tool descriptions tight and response payloads lean. Verbose tool schemas will burn your context budget fast.
Android ships today; iOS is coming. MCP server setup requires either a home machine or a cloud endpoint, which adds infrastructure overhead for prototyping.
Verdict: Evaluate. Worth prototyping if you're building local agents — but plan your MCP schema carefully from the start. Retrofitting concise tool descriptions after the fact is painful.
Node.js assigns CVE for all EOL versions
Node.js will now issue a single CVE identifier — flagged under CWE-1104 — covering all end-of-life versions. If your organization uses CVE tracking for dependency management, EOL Node.js versions will start triggering alerts in your security tooling.
This is a forcing function, not a new vulnerability. Node.js v16 is pulling 11 million downloads a month despite being EOL. The CVE mechanism exists to surface this through the pipelines organizations already trust rather than expecting developers to manually track EOL calendars. There is no workaround — the only resolution is migration.
Run node -v and npx is-my-node-vulnerable right now. If you're on v16, v14, or earlier, the upgrade targets are v18, v20, v22, or v23.
Verdict: Ship. This is not optional. If you're on EOL Node.js, the CVE will propagate through your scanners and the conversation will happen anyway — better to own the upgrade timeline than react to a scanner alert.
If this breakdown saved you some research time, Dev Signal publishes exactly this every issue — technically precise coverage of the AI developer tools that are actually moving. Worth subscribing if you'd rather read a sharp summary than wade through release notes.
Top comments (0)